CAD code generation on CADBench Sim
76.81Attr. FidelityGraph-CAD (SAPCL)
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Graph-CAD (SAPCL)Backbone=Qwen3-8B2026.04 | 76.81 | 74.23 | 55.46 | 68.83 | 2 | 66.93 | 90.18 | |
| Graph-CAD (SFT)Backbone=Qwen3-8B2026.04 | 72.95 | 72.65 | 47.33 | 64.31 | 2.4 | 65.44 | 78.3 | |
| Claude-opus-4-1inference_protocol=three-stage inference, reasoning_mode=true, shots=22026.04 | 72.16 | 73.68 | 54.03 | 66.62 | 7.4 | 61.51 | 49.32 | |
| GPT-5inference_protocol=three-stage inference, reasoning_mode=true, shots=22026.04 | 70.13 | 73.47 | 42.5 | 62.03 | 2.8 | 64.49 | 38.46 | |
| BlenderLLM2026.04 | 68.93 | 69.53 | 36.5 | 58.32 | 2.4 | 64.09 | 55.13 | |
| Deepseek-r1inference_protocol=three-stage inference, reasoning_mode=true, shots=22026.04 | 41.24 | 43.66 | 21.79 | 35.56 | 19.2 | 50.11 | 55.56 | |
| Qwen-Plusinference_protocol=three-stage inference, reasoning_mode=true, shots=22026.04 | 36.04 | 37.77 | 20.72 | 31.51 | 48.4 | 33.62 | 23.79 | |
| CADFusion2026.04 | 35.66 | 22.58 | 6.74 | 21.66 | 6.2 | 55.78 | — | |
| Llama-3.1-405binference_protocol=three-stage inference, shots=22026.04 | 33.02 | 33.55 | 15.37 | 27.31 | 36.4 | 39.43 | 32.69 | |
| Text2CAD2026.04 | 32.78 | 20.84 | 4.46 | 19.36 | 6.6 | 57.07 | — | |
| Gemini-2.5-proinference_protocol=three-stage inference, reasoning_mode=true, shots=22026.04 | 21.73 | 21.8 | 15.65 | 19.72 | 42.4 | 20.5 | 40.48 |